Synergistic spatial-sequential modeling for enhanced Rosa roxburghii disease detection.

The rapid advancement of smart agricultural technologies has enabled deep learning-based approaches to markedly enhance the efficiency of crop disease detection and quality monitoring. To overcome the challenges of low detection efficiency and high false-positive rates in large-scale cultivation of the characteristic economic crop Rosa roxburghii, this study introduces a synergistic spatial-sequential modeling framework. This synergy is achieved via a gated fusion mechanism that bidirectionally enhances CNN-extracted local spatial features and Mamba-derived global sequential features through adaptive re-weighting, forming a mutual refinement loop instead of a simple cascade. Built upon YOLOv8, the model integrates the Mamba module from state space models (SSMs) to develop a hybrid feature extraction framework, jointly optimizing detection accuracy and inference efficiency. A selective feature enhancement strategy further strengthens the network's ability to characterize subtle lesions on the surface of Rosa roxburghii by adaptively amplifying informative responses through channel-wise recalibration and spatial refinement. The backbone network also incorporates the selective scan for 2D data (SS2D) module to capture long-range dependencies beyond conventional convolution. Experimental validation conducted on our self-constructed dataset, which consists of field-captured images of Rosa roxburghii plants manually collected and annotated by the research team, demonstrates that the proposed model delivers notable performance gains, achieving mAP@50 improvements of 6%, 2.87%, 8.38%, 7.08%, 16.27%, and 6.55% over YOLOv5-N, YOLOv8-N, YOLOv10-N, YOLOv11-N, YOLOv12-N, and MambaYOLO-T, respectively. Concerning mean average precision (mAP)@ 50-95, the model attains corresponding gains of 6.85%, 0.27%, 6.99%, 5.19%, 13.66%, and 2.57%, all while maintaining comparable FLOPs (G) and parameter counts. These results highlight the model's excellent performance, confirming its effectiveness for agricultural visual inspection and suggesting a promising approach for intelligent disease monitoring in mountainous specialty crop production.